Video recommendations and Machine Learning

A good content recommendation system is key for any content provider. Machine Learning video recommendations provide a unique opportunity for broadcasters, Pay-TV operators, TV Networks, and any content distributor to increase engagement and reduce churn through content personalization.

Currently, most recommendation systems operate using explicit information provided by the user about their preferences (for example, by scoring previously watched content) using a technique known as collaborative filtering.

Most recommendation systems may also factor in a user’s profile when suggesting appropriate content, and in doing so may use the preference and feedback information provided by similar users. Such user profile information may contain demographic and geographic data, in addition to more dynamic data, such as the user’s web activity (e.g. pages visited, videos watched, activity on social networks).
Recommendation engines also use techniques that are based on the similarity of pieces of content, data that is used to make recommendations based on previously watched content.

Video Recommendation Technology

Today’s video recommendation technology uses machine learning to train, predict and provide video recommendations to video service users. For the most part, it uses algorithms to identify item similarity complemented with the user’s view history to produce a recommendation.
Item similarity: Users who liked this might also like…

If a user shows interest in specific content, similar content can be recommended via a non-personalised but effective recommendation algorithm known as a content-based-type recommendation.

The basic algorithm works as follows:
– To measure how similar two given items of content are, a feature vector, which encodes different scored metadata (E.g. genre information) should be assigned to each content and should compute the angle between each pair of vectors in Euclidean space. The smaller the angle between the vectors, the more similar the content is.
– Given an item of content, a shortlist of recommended similar items is produced by finding the most popular and best-rated examples among the most similar content.

The cold start issue: What to recommend to a new user?

While there is little user data about the likes and preferences of a new user, semi-personalised recommendations can be generated, based on demographic information (age, gender, country) together with genre-preferences submitted by the user.
The set of users is partitioned into predetermined clusters, and a given user is matched with a cluster. Within each cluster, a recommendation is produced by averaging the available ratings and producing a shortlist of the most popular, best-rated content.

Advanced machine learning techniques: In-content analysis:

Today new ways to measure content similarity are possible by using advanced machine learning techniques based on the recognition, analysis, and tagging of video frames and audio samples, which identify factors such as:
– Scene speed (fast, slow, loud, etc.)
– Colours and luminosity
– Scene locations (landscapes, city landmarks, indoor/outdoor scenes, etc.)
– Predominance of a time of day (morning, afternoon, night)
– Character sentiment and emotion detection (joy, sadness, violence, etc.)
– Predominant elements (sea, fire, sky, etc.)

This innovative approach uses neural networks and dimensionality reduction techniques and enables more advanced learning about user preferences based on a series of attributes obtained from previously watched audio-visual content. Not only can this increase recommendation accuracy but can, in turn, increase user engagement and reduce churn thereby impacting the overall ROI of the video service.

Do you want to know more? Please reach us HERE